Pith. sign in

REVIEW 2 major objections 5 minor 39 references

Hard physical constraints on flow matching let simulators force specific, realistic car-crash types on demand.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 19:03 UTC pith:QBSOZQIW

load-bearing objection Solid engineering transfer of hard-constraint flow matching into closed-loop AV scenario generation; large, ablated gains on public data, with feasibility gaps already quantified rather than hidden. the 2 major comments →

arxiv 2607.04451 v1 pith:QBSOZQIW submitted 2026-07-05 cs.CV

CCFM: Collision-Constrained Flow Matching for Safety-Critical Scenario Generation

Ke Li , Kaidi Liang , Yuxin Ding , Debojyoti Biswas , Xianbiao Hu , Ruwen Qin This is my paper
classification cs.CV
keywords safety-critical scenario generationflow matchinghard constraintsautonomous driving simulationGauss-Newton projectioncollision controllabilityclosed-loop evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Autonomous-vehicle planners need closed-loop tests against rare crashes, yet most generators only nudge trajectories with soft costs and cannot guarantee a rear-end, side, cut-in or head-on impact of chosen severity. This paper reframes the problem as collision-constrained sampling: a heuristic first picks the most plausible adversarial vehicle and crash type, then a flow-matching sampler is repeatedly projected onto three hard residuals (contact point, relative heading, closing speed) by a damped Gauss-Newton step. On two public driving datasets the method raises collision rates to 46–83 % while matching the requested crash type roughly 84 % of the time and keeping accelerations and off-road rates comparable to softer baselines. The result is a controllable, physics-grounded source of safety-critical scenes that can stress-test planners and supply synthetic crash data.

Core claim

Hard geometric and severity constraints can be enforced inside a learned flow-matching prior by intermediate Gauss-Newton manifold projections, producing controllable, type-specific collisions at far higher rates than soft-guidance methods while still yielding realistic multi-agent trajectories.

What carries the argument

Collision-Constrained Flow Matching (CCFM): at each ODE step an action sequence is projected onto the zero set of three residual functions (contact, heading, severity) via damped Gauss-Newton, then blended back with the optimal-transport interpolant so the terminal sample satisfies the desired crash type exactly.

Load-bearing premise

The intermediate projections stay feasible under the ego vehicle’s closed-loop reactions and do not push the trajectory so far off the learned traffic distribution that realism collapses.

What would settle it

Run the same closed-loop trials with a stronger ego planner that actively dodges; if type-match rate or collision rate falls sharply while residual infeasibility ratios rise, the hard-constraint claim fails under realistic interaction.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes Collision-Constrained Flow Matching (CCFM) for controllable safety-critical traffic scenario generation in closed-loop AV simulation. It reformulates the problem as constrained sampling from a learned flow-matching prior: a Heuristic Collision Selector (HCS) chooses an adversarial agent and one of four collision types (rear-end, side, cut-in, head-on); three residual constraints (contact point, relative heading, severity) define a feasible set at a dynamic time-to-collision T_col; and intermediate ODE samples are projected onto that set by damped Gauss–Newton with an OT reverse update (Alg. 1, Sec. 3.4). On nuScenes and nuPlan closed-loop rollouts against STRIVE, CCDiff and SAFE-SIM, CCFM reports substantially higher collision rates (up to 46.4 % / 83.1 % at 80-frame horizon; higher at 200 frames), ~84 % type-match, competitive realism, and planner-agnostic behavior, with ablations isolating HCS, each residual, and hard projection versus soft guidance.

Significance. If the empirical claims hold, the work supplies a practical route from soft cost guidance to hard geometric/severity control of named collision types while retaining a learned multi-agent prior. That combination is directly useful for closed-loop AV planner stress-testing and for generating labeled crash-like data. Strengths that raise confidence include public-dataset evaluation against fixed baselines, transparent composite scores, systematic ablations (Tables 5–7), a quantitative feasibility analysis (Supp. F.2), and released code. The transfer of hard-constraint flow matching from PDE settings to interactive traffic is a non-trivial methodological step for the community.

major comments (2)
  1. The central claim of “guarantees precise collision control through hard physical constraints” (abstract, Sec. 1) is stronger than the evidence. Supp. F.2 reports non-negligible infeasibility ratios (Inf_cnt up to 72.9 % for side, Inf_svt 39 % for head-on); Sec. 5.4 correctly notes that a collision is realized only when the constraint remains satisfiable under closed-loop ego reaction. The manuscript should restate the claim as empirical controllability under successive re-planning (CR 46–83 %, TM ~84 %) rather than a guarantee, and should surface the feasibility numbers in the main text so readers can judge coverage.
  2. Tables 1–7 report point estimates only (no standard errors, confidence intervals, or multi-seed statistics) for CR, TM, MS and composite scores. Given that HCS selection, stochastic flow sampling, and closed-loop ego reactions are all stochastic, the large reported gains need at least seed-level variability or bootstrap intervals before the magnitude of improvement can be treated as established.
minor comments (5)
  1. Free parameters of HCS (equal weights, geometry bandwidths/alphas in Supp. Table 2) and residual thresholds (Supp. Table 4) are stated but not sensitivity-tested; a short sensitivity paragraph or appendix table would strengthen reproducibility.
  2. Sec. 5.3 notes that avoidable vs. unavoidable collisions are not distinguished; even a simple post-hoc kinematic check would clarify planner relevance of the generated scenarios.
  3. Notation for residual maps ϕ_j and target values ℓ̃_j is split between main text and supplement; a compact main-text summary of the four residual triples would improve readability.
  4. Fig. 2 is qualitative only; adding a small quantitative panel (e.g., per-type TM from Supp. Table 5) next to the visualizations would make controllability more self-contained.
  5. Minor polish: arXiv id formatting, consistent hyphenation of “time-to-collision,” and ensuring all composite-score normalizations are defined before first use in tables.

Circularity Check

0 steps flagged

No significant circularity: hard residuals, GN projection, and empirical rates are independently defined and measured on external benchmarks.

full rationale

The paper defines four collision types via explicit residual functions (contact-point distance, relative heading, closing-speed severity) in Sec. 3.3 / Eqs. (9)–(10) and Supp. B, then enforces them by damped Gauss–Newton projection onto the feasible set C = {a | h_t = 0} inside the flow-matching ODE (Alg. 1, Sec. 3.4). These residuals are geometric/kinematic definitions, not fitted parameters. Collision rate, type-match, severity, and realism metrics are computed post-hoc from closed-loop rollouts on public nuScenes/nuPlan splits against fixed external baselines (STRIVE, SAFE-SIM, CCDiff); composite scores are transparent within-table normalizations. The projection idea is adapted from external PDE work (PCFM/HardFlow) and is not load-bearing via self-citation. Infeasibility ratios are reported rather than hidden (Supp. F.2). Nothing reduces a claimed prediction or first-principles result to its own inputs by construction; the elevated CR/TM figures are empirical outcomes of the constrained sampler under interactive ego reactions.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on a modest set of hand-chosen scoring weights, residual thresholds, projection hyper-parameters and the modeling assumption that unicycle dynamics plus a learned flow-matching prior remain valid after repeated hard projections. These are standard engineering choices rather than free parameters fitted to the final metrics; the invented entities are the four collision-type residual triples and the composite HCS score.

free parameters (5)
  • HCS score weights (w_rch, w_geo, w_lgt) = equal (1/3 each)
    Set equal by hand; sum to one. Directly determine which adversary and collision type are selected.
  • Geometry-score bandwidths and alphas (Table 2) = listed per type in Supp. Table 2
    Per-collision-type Gaussian widths and feature weights chosen to peak at desired relative poses.
  • Constraint thresholds (l̃_cnt, l̃_hdg, l̃_svt) and residual weights = e.g. rear-end severity ≥2 m/s, heading [0.95,1]
    Hard residual targets that define the feasible set C; listed in Supp. Table 4.
  • GN step size α and damping γ = α=0.8, γ=1e-4
    Control projection aggressiveness; fixed at 0.8 and 1e-4.
  • T_col clip range [T_min, T_max] = [5,10]
    Limits the dynamic time-to-collision window used for residual evaluation.
axioms (4)
  • domain assumption Vehicle motion obeys a deterministic unicycle dynamics model f that is differentiable for reverse-mode AD.
    Stated in Sec. 3.1 Eq. (1); required for both rollout and Jacobian computation inside the projection.
  • domain assumption A conditional flow-matching velocity field trained on naturalistic trajectories supplies a realistic action prior that remains useful after intermediate hard projections.
    Core modeling premise of Sec. 3.4; if the prior is destroyed by projection the realism claims fail.
  • ad hoc to paper The three residual functions (contact, heading, severity) evaluated at a single future horizon T_col are sufficient to encode the four named collision types.
    Defined in Sec. 3.3 and Supp. B; the paper’s controllability claim rests on this encoding.
  • domain assumption Closed-loop re-planning at 2 Hz with the chosen ego planners (IDM, PDM, lane-graph) is representative of realistic AV interaction.
    Experimental setup Sec. 4.2; results are conditioned on these planners.
invented entities (2)
  • Heuristic Collision Selector (HCS) composite score no independent evidence
    purpose: Dynamically choose the most threatening yet map-legal adversary–collision-type pair before constrained sampling.
    New scoring function combining reachability, geometry and legality kernels; no independent external validation beyond the paper’s own ablations.
  • Four collision-type residual triples (h_cnt, h_hdg, h_svt) no independent evidence
    purpose: Define the hard feasible set C that the Gauss-Newton projector enforces.
    Explicit mathematical encoding of rear-end/side/cut-in/head-on geometry and severity; invented for this work.

pith-pipeline@v1.1.0-grok45 · 26766 in / 3067 out tokens · 37898 ms · 2026-07-11T19:03:28.038262+00:00 · methodology

0 comments
read the original abstract

Evaluation of autonomous vehicle (AV) planners in safety-critical closed-loop simulation is essential for real-world deployment. However, generating controllable safety-critical scenarios remains challenging. Existing approaches use soft guidance that provides only probabilistic preferences and cannot guarantee the satisfaction of geometric and severity constraints associated with specific collision types. We introduce Collision-Constrained Flow Matching (CCFM), a novel framework that guarantees precise collision control through hard physical constraints. CCFM consists of three key components: (i) a heuristic collision selector that optimally identifies an adversarial agent and collision type via composite scoring; (ii) structured hard constraints that explicitly define four collision types (rear-end, side, cut-in, head-on) through contact point, heading, and severity requirements; and (iii) a collision-constrained flow matching sampler that enforces the constraints via Gauss-Newton manifold projection. CCFM achieves collision rate up to 46.4% on nuScenes and 83.1% on nuPlan, significantly outperforming baselines while preserving realistic driving behavior. By enabling controllable collision characteristics in safety-critical scenario generation, CCFM provides a reliable foundation for AV safety evaluation and sim-to-real crash data generation. The code and implementation details are available at https://github.com/KELISBU/CCFM.

Figures

Figures reproduced from arXiv: 2607.04451 by Debojyoti Biswas, Kaidi Liang, Ke Li, Ruwen Qin, Xianbiao Hu, Yuxin Ding.

Figure 1
Figure 1. Figure 1: Overview of the proposed safety-critical scenario generation by CCFM. 3.1 Safety-critical Traffic Generation in Closed-loop Simulation In the closed-loop simulation, we consider a traffic scene with N vehicles: an ego vehicle (i = ego) and N − 1 non-ego vehicles. One non-ego vehicle is selected as the adversarial agent and tasked with colliding with the ego vehicle, while the others are reactive agents tha… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative visualization of collision-type controllability. Given an assigned ego–adversary pair and scene context, CCFM generates distinct scenarios by specifying target collision types [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: Illustration of the geometry score Sgeo,t(i, m) for different collision types. For each collision type, the top row shows the target geometric configuration in terms of heading alignment di,t and bearing feature bi,t, while the bottom visualizes the corresponding geometry score landscape over the (di,t, bi,t) space. The star marks the target feature vector posm for each collision type m. The dot product di… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 5 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    2020

  2. [2]

    arXiv preprint arXiv:2106.11810 (2021)

    Caesar, H., Kabzan, J., Tan, K.S., Fong, W.K., Wolff, E., Lang, A., Fletcher, L., Beijbom, O., Omari, S.: nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810 (2021)

    Pith/arXiv arXiv 2021

  3. [3]

    In: Computer Vision – ECCV 2024

    Chang, W.J., Pittaluga, F., Tomizuka, M., Zhan, W., Chandraker, M.: SAFE-SIM: Safety-critical closed-loop traffic simulation with diffusion-controllable adversaries. In: Computer Vision – ECCV 2024. pp. 242–258 (2025)

    2024

  4. [4]

    In: 8th Annual Conference on Robot Learning (2024)

    Chen,K.,Lei,Y.,Cheng,H.,Wu,H.,Sun,W.,Zheng,S.:FREA:Feasibility-guided generation of safety-critical scenarios with reasonable adversariality. In: 8th Annual Conference on Robot Learning (2024)

    2024

  5. [5]

    In: Computer Vision – ECCV 2024

    Chitta, K., Dauner, D., Geiger, A.: SLEDGE: Synthesizing driving environments with generative models and rule-based traffic. In: Computer Vision – ECCV 2024. pp. 57–74 (2025) 16 K. Li et al

    2024

  6. [6]

    In: Conference on Robot Learning (CoRL) (2023)

    Dauner, D., Hallgarten, M., Geiger, A., Chitta, K.: Parting with misconceptions about learning-based vehicle motion planning. In: Conference on Robot Learning (CoRL) (2023)

    2023

  7. [7]

    IEEE Transactions on Intelligent Transportation Systems24(7), 6971–6988 (2023)

    Ding, W., Xu, C., Arief, M., Lin, H., Li, B., Zhao, D.: A survey on safety-critical driving scenario generation—a methodological perspective. IEEE Transactions on Intelligent Transportation Systems24(7), 6971–6988 (2023)

    2023

  8. [8]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA)

    Feng, L., Li, Q., Peng, Z., Tan, S., Zhou, B.: Trafficgen: Learning to generate diverse and realistic traffic scenarios. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3567–3575. IEEE (2023)

    2023

  9. [9]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., Alahi, A.: Social GAN: Socially acceptable trajectories with generative adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    2018

  10. [10]

    In: Computer Vision – ECCV 2022

    Hanselmann, N., Renz, K., Chitta, K., Bhattacharyya, A., Geiger, A.: KING: Gen- erating safety-critical driving scenarios for robust imitation via kinematics gradi- ents. In: Computer Vision – ECCV 2022. pp. 335–352 (2022)

    2022

  11. [11]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

    2016

  12. [12]

    In: Advances in Neural Information Processing Systems

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems. vol. 33, pp. 6840–6851 (2020)

    2020

  13. [13]

    In: International Conference on Machine Learning (2022)

    Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flex- ible behavior synthesis. In: International Conference on Machine Learning (2022)

    2022

  14. [14]

    In: 2018 IEEE Intelligent Vehicles Symposium (IV)

    Koren, M., Alsaif, S., Lee, R., Kochenderfer, M.J.: Adaptive stress testing for au- tonomous vehicles. In: 2018 IEEE Intelligent Vehicles Symposium (IV). pp. 1–7 (2018)

    2018

  15. [15]

    arXiv preprint arXiv:2511.08425 (2025)

    Li, Z., Alim, K., Azizan, N.: HardFlow: Hard-constrained sampling for flow- matching models via trajectory optimization. arXiv preprint arXiv:2511.08425 (2025)

    Pith/arXiv arXiv 2025

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Lin, H., Huang, X., Phan, T., Hayden, D., Zhang, H., Zhao, D., Srinivasa, S., Wolff, E., Chen, H.: Causal composition diffusion model for closed-loop traffic generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 27542–27552 (2025)

    2025

  17. [17]

    In: The Eleventh International Conference on Learning Representations (2023)

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  18. [18]

    Nature Communica- tions15(1), 4808 (2024)

    Liu, H.X., Feng, S.: Curse of rarity for autonomous vehicles. Nature Communica- tions15(1), 4808 (2024)

    2024

  19. [19]

    In: The Eleventh International Conference on Learning Representations (2023)

    Liu, X., Gong, C., qiang liu: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  20. [20]

    In: Advances in Neural Information Processing Systems

    Liu,Y.,Peng,Z.M.,Cui,X.,Zhou,B.:Adv-BMT:Bidirectionalmotiontransformer for safety-critical traffic scenario generation. In: Advances in Neural Information Processing Systems. vol. 38, pp. 55310–55335 (2025)

    2025

  21. [21]

    arXiv preprint arXiv:2509.20102 (2025)

    Nie, T., Mei, Y., Tang, Y., He, J., Sun, J., Shi, H., Ma, W., Sun, J.: Steerable adver- sarial scenario generation through test-time preference alignment. arXiv preprint arXiv:2509.20102 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Pang, B., Zhao, T., Xie, X., Wu, Y.N.: Trajectory prediction with latent belief energy-based model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11814–11824 (2021) CCFM 17

    2021

  23. [23]

    arXiv preprint arXiv:2505.11247 (2025)

    Peng, M., Xie, Y., Guo, X., Yao, R., Yang, H., Ma, J.: Ld-scene: Llm-guided diffusion for controllable generation of adversarial safety-critical driving scenarios. arXiv preprint arXiv:2505.11247 (2025)

    Pith/arXiv arXiv 2025

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Rempe, D., Philion, J., Guibas, L.J., Fidler, S., Litany, O.: Generating useful accident-prone driving scenarios via a learned traffic prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17305–17315 (2022)

    2022

  25. [25]

    In: Computer Vision – ECCV 2020

    Salzmann, T., Ivanovic, B., Chakravarty, P., Pavone, M.: Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data. In: Computer Vision – ECCV 2020. pp. 683–700 (2020)

    2020

  26. [26]

    In: Interna- tional Conference on Learning Representations (2021)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: Interna- tional Conference on Learning Representations (2021)

    2021

  27. [27]

    38310–38335 (2025)

    Tan, T., Zheng, Y., Liang, R., Wang, Z., Zheng, K., Zheng, J., Li, J., Zhan, X., Liu, J.: Flow matching-based autonomous driving planning with advanced interactive behaviormodeling.In:AdvancesinNeuralInformationProcessingSystems.vol.38, pp. 38310–38335 (2025)

    2025

  28. [28]

    arXiv preprint arXiv:2302.00482 (2023)

    Tong, A., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Fatras, K., Wolf, G., Bengio, Y.: Conditional flow matching: Simulation-free dynamic optimal trans- port. arXiv preprint arXiv:2302.00482 (2023)

    Pith/arXiv arXiv 2023

  29. [29]

    Physical review E62(2), 1805 (2000)

    Treiber, M., Hennecke, A., Helbing, D.: Congested traffic states in empirical ob- servations and microscopic simulations. Physical review E62(2), 1805 (2000)

    2000

  30. [30]

    In: The Thirty-ninth Annual Conference on Neural Information Process- ing Systems (2025)

    Utkarsh, U., Cai, P., Edelman, A., Gomez-Bombarelli, R., Rackauckas, C.V.: Physics-constrained flow matching: Sampling generative models with hard con- straints. In: The Thirty-ninth Annual Conference on Neural Information Process- ing Systems (2025)

    2025

  31. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wang, J., Pun, A., Tu, J., Manivasagam, S., Sadat, A., Casas, S., Ren, M., Ur- tasun, R.: AdvSim: Generating safety-critical scenarios for self-driving vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9909–9918 (2021)

    2021

  32. [32]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Xing, Z., Zhang, X., Hu, Y., Jiang, B., He, T., Zhang, Q., Long, X., Yin, W.: GoalFlow: Goal-driven flow matching for multimodal trajectories generation in end-to-end autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1602–1611 (2025)

    2025

  33. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence39(8), 8797–8805 (2025)

    Xu, C., Petiushko, A., Zhao, D., Li, B.: DiffScene: Diffusion-based safety-critical scenario generation for autonomous vehicles. Proceedings of the AAAI Conference on Artificial Intelligence39(8), 8797–8805 (2025)

    2025

  34. [34]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA)

    Xu, D., Chen, Y., Ivanovic, B., Pavone, M.: Bits: Bi-level imitation for traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2929–2936. IEEE (2023)

    2023

  35. [35]

    In: Computer Vision – ECCV 2024 Workshops

    Yin, Y., Khayatan, P., Zablocki, É., Boulch, A., Cord, M.: ReGentS: Real-world safety-critical driving scenario generation made stable. In: Computer Vision – ECCV 2024 Workshops. pp. 262–276 (2025)

    2024

  36. [36]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zhang, J., Xu, C., Li, B.: ChatScene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15459–15469 (2024)

    2024

  37. [37]

    In: Proceedings of The 7th Conference on Robot Learning

    Zhang, L., Peng, Z., Li, Q., Zhou, B.: CAT: Closed-loop adversarial training for safe end-to-end driving. In: Proceedings of The 7th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 229, pp. 2357–2372 (2023) 18 K. Li et al

    2023

  38. [38]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA)

    Zhong, Z., Rempe, D., Xu, D., Chen, Y., Veer, S., Che, T., Ray, B., Pavone, M.: Guided conditional diffusion for controllable traffic simulation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3560–3566 (2023)

    2023

  39. [39]

    αm 1−α m #T

    Zhou, Z., Hu, H., Chen, X., Wang, J., Guan, N., Wu, K., Li, Y.H., Huang, Y.K., Xue, C.J.: BehaviorGPT: Smart agent simulation for autonomous driving with next-patch prediction. In: Advances in Neural Information Processing Systems. vol. 37, pp. 79597–79617 (2024) Supplementary Material: CCFM: Collision-Constrained Flow Matching for Safety-Critical Scenari...

    2024